Reading Minds and Leaked LLaMAs This Week in ChatGPT

Author

The AI Bloke
March 09, 2023

You aren't ready for the amount of AI heading your direction... unless your name is Sam Altman. Then this may be a little underwhelming. But only a little. 

Welcome to The Week in ChatGPT newsletter where we discuss groundbreaking products powered by ChatGPT and the latest news going around the AI community. If you have any products or tidbits that you would like to be featured, by all means please drop them here. If your friend is forwarding you this email and you have no idea what is going on subscribe here, so you can keep getting confused every week! Onwards!

Before we get to any of the news this week, something of extreme importance has come to my attention. Our newsletter mascot (seen below) still does not have a name! A grave oversight on my part. Please click here and comment any name you think fits. The name that gets the most likes wins and whoever has the winning name will get a shoutout on the newsletter. Also, while you're there, please follow me so I can badger you on social media in addition to email. 

News Headlines

  • AI Can Now Read Your Mind 🫠

  • LLaMA LEAKS 😱

  • Quora Enters the AI Fray

Bodacious AI Products — Fun Edition

✈️ An AI Flyby for the Week ✈️

I'm not even joking. Researchers from the Osaka University have showed that it's possible to reconstruct images in peoples' heads by looking at fMRI scans and then feeding the observed features to an image generation model like Stable Diffusion. The researchers are essentially mapping regions of brain activity to a text and image encoder (upper brain region for text and lower brain regions for vision) and can use this information to recreate semi-accurate representations of what the individual was originally looking at. This was done without any model fine tuning for individuals, meaning it worked for different brains of different people without any modifications. Here's an example of what it can do!

You can see the image reconstructions are by no means exact, but this still might be the coolest application of diffusion models I've ever heard of. I wonder if they will do a follow up study and have people imagine pictures without looking at a presented image and see if they can still reconstruct images that are in peoples minds. This would literally enable us to create and share images straight from our brains, meaning we wouldn't even need to type out prompts anymore. We would just need to think it and it would be projected on the screen for people to see 😨. Forget, text to image. I'm bullish on brain to image. Absolutely crazy stuff. READ MORE HERE

If you don't know what a rickroll is, click here to find out. 

Remember how last week Zuck and Meta released LLaMA, a set of new large language models that serve as a ChatGPT competitor. Well, LLaMA has been leaked. It was supposed to be available specifically for research institutions and other organizations that Meta deemed worthy to prevent any misuse, but looks like that plan went up in smoke pretty much immediately.

The weights for the model showed up on 4chan and Torrent so now anybody who has 100 GB of RAM lying around can train their own chatbot (not me I'm barely making do when I open up Chrome). Now, you too can enjoy the beefy LLM that is LLaMA for the small (not small) price of hundreds of dollars of cloud storage READ MORE HERE.

Quora wants to throw hands in the AI ring too. It recently came out with Poe, which lets people "ask questions, get instant answers, and have back-and-forth conversations with several AI-powered bots". Poe provides access to 4 popular chatbot services, including ChatGPT, Anthropic’s Claude, Sage, and Dragonfly. Basically, if you don't use Quora for Quora, you can always use it to get access to ChatGPT 😂 or any of these other models. READ MORE HERE

Now For Some AI Products 

PRODUCT 1: ScribbleDiffusion — Scribble With Style

Have you every wished you were an artist. Well, with the advent and popularization of models like DALLE and Midjourney, people are able to create images from descriptions, turning artistry into a literary task. However, recently, another alternative was released called ScribbleDiffusion. It enables people to augment their sketches and scribbles into higher quality images with text. Here's a look at the interface. 

Look, guys, I never claimed to be an artist so please cut me some slack and just be glad I didn't attempt a self-portrait. That could have been blindingly bad. Yet despite how bad my initial sketch was, the generated photo based on my drawing plus the context I provided of "blue shark" transformed my subpar sketch into a legitimately nice image. You can see that the shark resembles mine in shape and positioning, but has been improved to the point of looking realistic.

For all those with any artistic inclinations, I would be interested to see what kind of images are possible for sketches that don't look like poop. Put 'em up on Twitter and don't forget to tag @TheAIBloke. 

PRODUCT 2: Roastedby.ai  — My Feelings Hurt

I can say with total honesty and some sadness that Roastedby.ai is good at its job. I tried to out-roast it, but got grilled like a kabob repeatedly. How do you even successfully roast a computer program?  Maybe something like "Yo code so fat, I bet it got a lot of bugs" 😂. Well, anyway, here's a screenshot of one of many conversations that I unfortunately engaged in with this software. 

To summarize, I got called dull as a spoon and then the AI hit me with "Have you considered a career in modeling? I hear there's high demand for potato sacks". Internally, I was very distraught, but tried to play it off and asked why they would want potato sacks as models. I should not have asked because the AI had a response for that locked and loaded: "Well, I thought you'd be perfect for it, considering how much you resemble a lumpy, misshapen potato. But don't worry even potatoes can have some appeal". Essentially, if you every want to experience impotent rage for no particular reason or perhaps just want to trade insults with a computer, this is the perfect site for you.

Post any highlights of your conversation with our resident roaster on Twitter and tag me so we can shit-talk on Twitter, where we are safe from its vicious comebacks. 

Where My Reverse Engineers At?

I feel like this image did not get the love it deserved, so I'm reusing it for this newsletter. 

See if you can guess the prompt that was used to generate the image below (try and reverse engineer the prompt from picture in engineer parlance).

Whoever has the closest guess earns my undying respect and 5 bucks (if they have a crypto wallet). Here's the thread on my profile and comment whatever you think the seed prompt was for the image. If you don't want to do the challenge and just want to make some cool pictures, go ahead and Tweet those too! But I doubt they're gonna be cooler than these little guys! Also if nobody participates, the AI Bloke will be very sad. 

Here's the images for this week (HINT: GENERATED WITH 2 WORDS ): 

Let the guesses begin! I don't know how, but somehow this was initially created with just two words unless the Midjourney Discord is straight up lying to me (Midjourney creates 4 variations for any default prompt). Basically, generative image models = very cool. 

Parting is Such Sweet Sorrow (But Please Keep Reading LOL)

Well that's a wrap on the newsletter.

Let me know how I did by filling out this feedback form. If I did well and you're in the giving spirit please support the newsletter by buying me a coffee or donating to my gumroad page. Most of the money you donate to me will go back into purchasing AI products to try out and review on the newsletter. Bernie says it better:

Also if you have a product that you would like to be featured on the newsletter remember to add it here! 

 Lastly, if you'd like to refer a friend just tell them to visit chatgptnewsletter.com to subscribe! Or you could always just forward them the email without any subject header. I'm sure they they would love that.

Thank you and see you all next week!

Cool Things You Should Try/Buy

(Newsletter 1) ==> Consult domain experts in your browser with ExperAI

(Newsletter 3)  ==> Pay for DoNotPay lol. Irony aside, it really is pretty cool.

(Newsletter 4)  ==> Multion.AI - not your average burger buying extension. 

(Newsletter 5)  ==> Why settle for regular email when you could have Intellimail?   

(Newsletter 6)  ==> The AI app store where you can Cookup just about anything

(Newsletter 7)  ==> A cute, fuzzy AI-powered online meeting summarizer: Otter.ai

(Newsletter 7)  ==> Tired of taking hours to find the perfect online purchase? Getproduct.help's got you covered

Reply

or to participate.